Texture bags: anomaly retrieval in medical images based on local 3d-texture similarity

Providing efficient access to the huge amounts of existing medical imaging data is a highly relevant but challenging problem. In this paper, we present an effective method for content-based image retrieval (CBIR) of anomalies in medical imaging data, based on similarity of local 3D texture. During learning, a texture vocabulary is obtained from training data in an unsupervised fashion by extracting the dominant structure of texture descriptors. It is based on a 3D extension of the Local Binary Pattern operator (LBP), and captures texture properties via descriptor histograms of supervoxels, or texture bags. For retrieval, our method computes a texture histogram of a query region marked by a physician, and searches for similar bags via diffusion distance. The retrieval result is a ranked list of cases based on the occurrence of regions with similar local texture structure. Experiments show that the proposed local texture retrieval approach outperforms analogous global similarity measures.